Guide · AI security · as of 12 November 2026

Your AI server trusts localhost. Attackers are now exploiting that. Four checks.

On 9 November 2026 the US cybersecurity agency CISA added a flaw in GPUStack to its list of vulnerabilities known to be exploited in the wild. GPUStack is open-source software that manages GPU machines and serves AI models, the kind of thing a company installs to run AI on its own hardware. The flaw is not exotic. The software gave full administrator rights to any request that looked like it came from the same machine, and an outsider could make a request look that way. This guide works from the official record and turns it into four checks for any AI server you host.

What the record says

Same machine meant administrator.

The National Vulnerability Database entry for CVE-2026-102895 says that before version 2.2.0, GPUStack treated a request from the loopback address as the built-in administrator whenever the setting force_auth_localhost was off, which was the default. A remote request could appear to come from 127.0.0.1 in two ways: through a reverse proxy running on the same host, or through a header the client supplies itself, X-GPUStack-Real-IP. The entry says the bypass allowed unauthenticated administrative access. It is scored 9.8 out of 10.

What an attacker could do

Models, keys, secrets, and code.

The record lists what the bypass allowed: changing users and models, creating administrator API keys that keep working, reading worker and cloud-provider secrets, and running code through one of the server’s own endpoints. It also says exploitation was observed in the wild against version 2.1.1, and that version 2.2.0 fixes the flaw. For a business, the secrets are the costly part. An AI server usually holds keys to your cloud account, your storage and the models you pay for. Whoever controls the server holds those keys too.

Four checks for any AI server you host

Do not let “local” mean “trusted”.

  • 1. Update the software. If you run GPUStack, move to 2.2.0 or later. Test: read the version in the admin page, then compare it with the vendor’s release notes.
  • 2. Require a login for every request, including local ones. Turn on any setting that forces authentication on localhost. Test: call the API from the server itself without a key and confirm it is refused.
  • 3. Keep the server off the open internet. Put it behind a VPN or an identity-aware proxy, and bind it to a private address. Test: from a network outside yours, try to reach its port.
  • 4. Assume the keys it holds will leak one day. Give the server cloud keys that can do one job, with spend limits, and rotate them after any exposure. Test: list every key on the server and what each one can do.

What it means for a business

Private AI is only private if the door is locked.

Running AI on your own hardware keeps your data out of someone else’s cloud, and that is a real benefit for UAE firms with confidential records. It also makes you the operator, so the patching and the access rules become your job. Many AI tools were built on a laptop first, where everything on the machine can be trusted. When the same tool moves onto a shared server, that habit is the weak spot. Ask of every AI service you host: who can reach it, and does it ask them to prove who they are?

Private AI in the UAE, on your own server

Questions

Questions about the GPUStack flaw.

  • What is GPUStack?

    The NVD describes it as an open-source GPU cluster manager for AI model serving and GPU instance provisioning. Companies use tools like it to run AI models on their own graphics cards.

  • Am I affected?

    Only if you run GPUStack at a version before 2.2.0. If you do, update first, then check whether anyone could have reached the server and rotate the secrets it held. If you do not run GPUStack, the lesson still applies to any AI server.

  • What does “known exploited” mean?

    CISA keeps a catalogue of vulnerabilities it has evidence are being used in real attacks. A flaw on that list should be fixed before ordinary updates. CISA added this one on 9 November 2026.

  • Is running AI on our own servers a bad idea?

    No. It is often the right choice when data must stay in the country or in your hands. It needs the same care as any server: updates, a login on every request, and no direct exposure to the internet.

Sources

read 12 November 2026

Check it at the source.

  1. NVD: CVE-2026-102895, GPUStack authentication bypass (the description, the score and the fixed version)
  2. GPUStack: security advisory GHSA-gj9m-f8ph-pmhr (the maintainers’ own advisory)
  3. GPUStack: release 2.2.0 (the fixed version)

A question?

Run AI on your own hardware?

Tell us which AI services your team hosts and who can reach them. We will check the doors, the updates and the keys with you.

Talk to usWhatsApp +971 50 728 5494hello@brevdigital.com